Cursor Router¶

Development line: project:router · thread router
Last event: 2026-07-25 · 1 dated since 2026-07-25 · Researched: 2026-09-05 · confidence: high

What it is¶

Cursor Router routes each Agent request across Cursor’s changing model pool instead of making a developer choose a model per request. It offers Cost, Balance, and Intelligence modes for Teams and Enterprise. Balance and Intelligence use usage limits faster than Cost, so pick an explicit mode rather than treating Auto as a fixed model. The router suits team-level cost and quality control, but not workloads that require a pinned underlying model.

Development line¶

• 2026-07-25 — Cursor published a public update associated with Router. On 2026-07-25, Cursor published a public blog article associated with Router. The post marks a public milestone for the Cursor Router development line, but Cursor did not detail its functionality, rollout status, or user impact.

What changed¶

• 2026-07-25 — Cursor Router replaced Auto’s fixed routing with per-request classification and introduced Cost, Balance, and Intelligence controls for Teams and Enterprise.
• 2026-08-06 — Cursor documented the learned two-stage routing approach and reported post-launch improvements from more production traffic.
• 2026-08-12 — Cursor published rollout guidance: enable it for a busy team first, monitor cost per commit, then widen only if the measured result holds.

How to use this¶

From 2026-07-25, consult Cursor's Router announcement before relying on or configuring Router.

1. In Cursor’s model picker, select Auto, then choose Cost for spend control, Balance for the default compromise, or Intelligence for harder multi-step work. — https://prod.cursor.com/docs/cursor-router
2. For an Enterprise team, enable Router in the team dashboard; configure allowed modes, model access, visibility of the routed model, and optional soft or hard Auto enforcement. — https://prod.cursor.com/docs/cursor-router
3. In the TypeScript or Python Agent SDK, use model ID auto-smart and optimize_for set to cost, balanced, or intelligence; call Cursor.models.list() before relying on it for a team. — https://prod.cursor.com/docs/cursor-router

Best practices¶

• Pilot Router with one active team and measure cost per commit for a week before expanding the rollout. — https://cursor.com/guides/model-routing
• Keep a price-efficient routing option enabled; Cursor’s Enterprise documentation says blocking too many models can reduce quality or disable Router, and names Cursor Grok 4.5 as required for Router to work. — https://prod.cursor.com/docs/cursor-router
• Do not guess which model answered from output alone: the routed model is hidden by default, and the pool can change as models ship. — https://prod.cursor.com/docs/cursor-router

Superseded by this¶

• 2026-07-22 — treating Auto as one stable, manually predictable model choice is obsolete for routed Balance and Intelligence requests; Cursor now chooses the underlying model per request.
• 2026-07-22 — the earlier Auto behavior remains only as Cost mode, according to current documentation.

Still unknown¶

• Cursor does not publish a stable, exhaustive mapping from request types to underlying model names; the model pool changes over time.
• The published savings and satisfaction figures are first-party measurements, not an independently reproducible benchmark.
• Specific workflow changes remain unverified until the announcement article is researched.

Sources¶

Agent brief¶

• Subject: project:router, thread router, 1 dated events 2026-07-25 → 2026-07-25.
• Practical note: From 2026-07-25, consult Cursor's Router announcement before relying on or configuring Router.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.